I'm a passionate Applied Mathematics & Computer Science student at NJIT with a focus on AI/ML and data science. Currently conducting research on Beyond Next-Token Prediction paradigms for Large Language Models, while leading impactful community organizations.

current_focus:
  - Novel LLM architectures (intent-inference over next-token prediction)
  - Advanced transformer architectures & Graph-RAG systems
  - Model Context Protocol (MCP) implementation

collaboration:
  - AI/ML projects
  - Hackathons & competitions
  - Community tech initiatives

achievements:
  - Raised $70K for student organization
  - Led 100+ students to national conferences
  - 2x increase in student internship placements

• Advanced MLOps & Model Deployment
• Graph Neural Networks (GNNs)
• Quantum Computing Applications
• Agentic AI Systems

What it does: A controlled testbed for subliminal learning — the claim that a student distilled on a teacher's outputs over semantically unrelated data still inherits the teacher's trait. Reproduces the effect on MNIST, then probes the mechanism.

Verified baseline (M = 25 seeds):

• 🎯 Teacher accuracy 0.944, chance 0.10
• ✅ Same-init student on pure noise: 0.222
• ❌ Cross-init student on pure noise: 0.122
• 📐 Weight-update alignment η_cos: 0.254 vs 0.029

Three pre-registered hypotheses:

• H1 — tangent-space geometry vs. semantic overlap
• H2 — smooth continuum vs. hard threshold as inits diverge
• H3 — bandwidth ceiling vs. sample complexity

The honest headline:

None of the three favored hypotheses earned a clean confirmation — and that is the finding. H1's pre-registered falsifier was actually met, pointing away from the geometry story; the remaining confound is flagged in the writeup rather than buried.

Predictions committed to docs/PREDICTIONS.md before any run, plus a 31-item adversarial code-verification pass.

# Distillation on noise, not MNIST
out = student(bx)[:, :, idx]          # [M,B,L]
with t.no_grad():
    tgt = teacher(bx)[:, :, idx] * teacher_temp

if mode == "soft":
    loss = F.kl_div(
        F.log_softmax(out, -1),
        F.softmax(tgt, -1),
        reduction="batchmean",
    )
# bx is random noise — the student never
# sees a single real digit.

Shared architecture:

mlp:      [784, 256, 256, 13]
logits:   10 real + 3 ghost
seeds:    25 (for CIs)
device:   mps -> cuda -> cpu

Tech Stack:

📄 Ships FINALPAPER.pdf, an executive summary, and a literature review.

Evaluation (own rubric):

The negative result:

finding: >
  DPO gave no improvement over SFT
  due to preference saturation —
  chosen and rejected samples came
  from the same model family.
status: reported, not hidden

Tech Stack:

What it does: Fine-tunes Qwen 2.5 7B-Instruct with LoRA (via MLX on Apple Silicon) to answer philosophy and theology questions in a scholastic, Latin-inflected register with citations — a study in register transfer, not a doctrinal authority.

The pipeline:

scrape primary texts  →  clean
   ↓  (Catechism, Summa Theologica,
       Confessions, City of God)
generate pairs w/ Claude as teacher
   ↓
convert to MLX Q8  →  LoRA SFT
   ↓
rubric eval  →  optional DPO

What shipped:

• 🤗 Four LoRA adapters published on Hugging Face
• 🎛️ Live Gradio demo on a Hugging Face Space
• 📄 arXiv-style paper and conference poster, auto-rendered by GitHub Actions on every push
• 🔬 A post-experiments/ directory of follow-up probes (content-test, layer-frame, math-test), each with its own FINDINGS.md

Status: Complete through Phase 2 (scaled SFT + DPO experiment).

What it does: A from-scratch agent harness. A local daemon runs the full context → model → stream → tools loop and exposes it over a versioned JSON-RPC 2.0 / WebSocket contract; the terminal UI and browser UI are just two clients of that same protocol.

Design decisions worth defending:

• 📜 The wire contract (schema/wire-v0.json) was frozen up front — and did not change once across the entire build
• 🔌 Ollama, OpenAI, and Anthropic backends implemented directly on httpx — no vendor SDKs
• 🧱 import-linter enforces ports/adapters in CI: protocol and ports are forbidden from importing any vendor package
• 🎯 Retrieval-based skill selection scopes each turn to a subset of tools before the model is called (falling back to the full set if scoping would unlock nothing)
• 💾 Sessions persist to SQLite (WAL) + per-session JSONL transcripts, and replay across a daemon restart

Quality gates: ruff, pyright --strict, import-linter, and pytest all run in CI.

Status: self-labeled v0.1 — shipped. A working prototype with an architecture designed to be extended, not a maintained product.

# The agent turn loop, capped at 8 iters
async def run_agent_turn(
    adapter, registry, messages, emit, *,
    selector=None,
):
    tools = registry.specs() if registry else None
    if selector and tools:
        messages, tools = await _apply_skill_selection(
            selector, messages, tools
        )
    for _ in range(_MAX_TOOL_ITERS):  # = 8
        async for ev in adapter.stream(messages, tools):
            if isinstance(ev, TextChunk):
                await emit(TokenEvent(text=ev.text))
            elif isinstance(ev, ToolCallRequest):
                tool_calls.append(ev)
        if not tool_calls:
            await emit(DoneEvent(reason="complete"))
            return

Tech Stack:

% Every chapter, same contract:
\begin{definition} ... \end{definition}
\begin{theorem}    ... \end{theorem}
\begin{proof}      ... \end{proof}
% ...then runnable code.

The 9 blocks:

01. Foundations       # sets, proofs
02. Probability & Info
03. Stochastic Optimization
04. Neural Networks
05. Attention & Transformers
06. Pre-training
07. Post-training     # SFT, RLHF, DPO
08. RL                # MDPs -> GRPO
09. Inference & Serving

Tech Stack:

What it does: A 34-chapter derivation chain that starts at set theory and does not stop until it reaches KV caching and speculative decoding. Every claim is derived, not described.

What makes it unusual:

• 📚 Every chapter exists in three synchronized forms — Markdown, LaTeX/PDF, and Jupyter — with generate.py verifying all three stay in sync and regenerating TOCs
• 🤖 CI renders the PDF and HTML automatically on every push
• 🔨 Chapter 27 ships two GPT pre-training implementations: a pure-NumPy no-dependency baseline (~18K params) and a real PyTorch/MLX GPT (~30M params, 6 layers, d_model=384, tiktoken GPT-2 vocab) on TinyStories
• 🧮 Recent work has focused on closing logical gaps between chapters — filling in Borel σ-algebra existence, Pinsker's inequality, the performance-difference lemma

Companion repo: math-foundations holds the atomic math reference pages this chain cross-links into.

⚠️ The chapter-27 hardware benchmark table is still a template — the training run is implemented, but the wall-clock and throughput numbers have not been measured yet.

Society of Hispanic Professional Engineers

team_size: 20+ officers
member_base: 300+ students
funding_raised: $70,000+
conference_attendees: 100+ students

Key Achievements:

• 🤖 Built CLARA - AI Assistant for operations
  • Email management & auto-response
  • Event planning & logistics
  • Member engagement tracking
• 📈 2x increase in student internship placements
• 🎯 Led 100+ students to SHPE National Conference
• 🤝 Established partnerships with 15+ tech companies

Technologies Used: Python LangChain GPT-4 Gmail API Google Calendar API

Association of Latino Professionals For America

status: First ALPFA chapter at NJIT
focus: Non-engineering Latino professionals
partnerships: 5+ local ALPFA chapters

Key Achievements:

• 🏢 Established first professional Latino org for non-engineers
• 🌐 Built external partnerships network
  • Connected with NYC ALPFA chapters
  • Formed corporate partnerships
• 📚 Organized professional development workshops
• 🎤 Hosted industry speaker series
• 💼 Created mentorship program matching students with professionals

Impact: Expanded professional opportunities beyond engineering, creating an inclusive community for all Latino students.

Beyond Next-Token Prediction

Currently exploring novel paradigms for Large Language Models that reframe training from simple next-token prediction toward intent-inference and reasoning capabilities.

Key Areas:

• 🧠 Intent-based learning architectures
• 🔄 Interactive reasoning systems
• 📊 Multi-modal transformers
• 🎯 Context-aware embeddings

Building the Future of AI

• 🤖 CLARA 2.0: Enhanced AI assistant with advanced reasoning
• 📊 Graph-RAG Framework: Novel retrieval architecture for LLMs
• 🔗 MCP Implementations: Model Context Protocol integrations
• 🎓 ML Education Platform: Teaching AI/ML to fellow students